Thanks, and great question - we think about this a lot and think there are a couple of things here.

First, as models get better, our agent's ability to navigate a website and generate accurate automation scripts will improve, giving us the ability to more confidently perform multi-step generations and get better at one-shotting automations.

We expect browser agents will improve as well, which I think is more along the lines of what you're asking. At scale, we still think scripts will be better for their cost, performance, and debuggability aspects - but there are places where we think browser agents could potentially fit as an add-on to deterministic workflows (e.g., handling inconsistent elements like pop-ups or modals). That said, if we do end up introducing a browser agent in the execution runtime, we want to be very opinionated about how it can be used, since our product is primarily focused on deterministic scripting.

> At scale, we still think scripts will be better for their cost, performance, and debuggability aspects

This actually makes a ton of sense to me in lots of the LLM contexts (e.g. seeing how we are starting to prefer having LLMs write one-off scripts to do API calls rather than just pointing them at problems and having them try it directly).

Thanks!